Research map / Point clouds / Shape and surface completion

Geometry-Centered 3D Latent World Models for Growing Surfaces

NeurIPS 2026 · Point clouds › Shape and surface completion · approach: Point cloud diffusion generation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FOLIAGE, a geometry-centered latent world model for growing thin surfaces that allocates a fixed state budget across mature scaffold, growth-front, and interface tokens, with privileged physical energies used only at training, and introduces the SURF-GARDEN/SURF-BENCH simulated benchmark.

Problem

World models for growing surfaces must infer a current geometric state from partial heterogeneous observations and predict futures whose support (material and geometry) expands over time under material controls, with hidden material response unobserved at deployment.

Approach

Growth-adaptive state allocation partitions a fixed token budget into mature scaffold, active growth-front (high-resolution), and interface tokens, with role-specific dynamics; material actions enter only in the transition, and per-vertex physical energies supervise training via an EMA target encoder and activity supervision, not at deployment.

Key results

• Inverse-material MAE 32 (x10^3) vs 53-60 for best baselines (~40% reduction); 5-step mesh forecasting Chamfer 26 vs 38 for best baseline.
• Cross-modal retrieval mAP@100 64/39/37 vs 48/22/23 for PointCLIPv2, a +14 point gain in the supervised image-to-point setting.
• Passive real plant forecasting Chamfer (x10^3) 19.6 on Pheno4D Maize vs 23.1 for Uniform AR light-finetuned.

Surprising findings

• Placing the action in the observation encoder reduces branch retrieval from 90.4% to 78.2%, supporting the action-in-transition interface.
• Under matched privileged supervision, FOLIAGE's forecasting advantage persists (C40 41.2 vs 62.7 for uniform allocation).

Limitations

• Scoped to accretive thin surfaces; excludes volumetric growth, tearing, and merging.
• Degrades under globally distributed activity, poor sensing, and long open-loop rollouts.
• Counterfactual intervention evaluation is simulation-only; real data lacks matched interventional futures.

Experimental setup

Datasets: SURF-GARDEN GrowliFlower Pheno4D Crops3D D-FAUST CAPE

Benchmarks: SURF-BENCH GrowliFlower Pheno4D Crops3D D-FAUST CAPE

Metrics: accuracy geodesic error mAP@100 MAE balanced accuracy Chamfer distance vertex drift branch retrieval separation ratio

Baselines: MeshCNN DiffusionNet ETNN ZoomOut G-MSM CrossPoint CLIP2Point ULIP-2 PointCLIPv2 DiffPD BDP DiffCloth

Builds on: DiffusionNet PointNeXt GRAND++ DiffusionNet-based correspondence

Closest papers by meaning

• Single Mesh Diffusion Models with Field Latents for Texture Generation (CVPR 2024 · similarity 0.85)
• PhysX-3D: Physical-Grounded 3D Asset Generation (NeurIPS 2025 · similarity 0.85)
• GALA: Geometry-Aware Local Adaptive Grids for Detailed 3D Generation (ICLR 2025 · similarity 0.85)
• SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · similarity 0.85)
• Geometric Trajectory Diffusion Models (NeurIPS 2024 · similarity 0.84)
• CraftsMan3D: High-fidelity Mesh Generation with 3D Native Diffusion and Interactive Geometry Refiner (CVPR 2025 · similarity 0.84)